One of the pioneers in the field of agent-based simulation of the electricity sector is a research team at London Business School. Day and Bunn (Day and Bunn, 2001) describe a simulation where generation companies bid their individual piece-wise linear supply function in a market with uniform price market clearing. The generation companies are modelled as daily profit maximisers who assume that the competitors bid the same supply function as they did in the previous day. In addition to the profits obtained from electricity sales, companies can also earn revenue from so-called contracts for differences, i.e. financial hedging contracts with buyers of electricity. Each company selects its best action through an iterative optimization routine that calculates the supply function that the company bids into the pool market with
respect to the mentioned conjecture about the opponent's actions and a given volume of con-tracts for differences. Electricity demand is represented by an aggregate demand function with a defined slope. Day and Bunn evaluate their model by comparing the supply functions ob-tained from the computational model with the equilibrium in continuous supply functions that would be obtained through the approach formulated by Klemperer and Meyer (Klemperer and Meyer, 1989). A central finding is that the results are reassuringly close in the studied sce-nario, which models competition between the three largest fossil fuel generating companies in the England and Wales pool. They conclude that the computational approach can also deliver realistic results for more complex scenarios that cannot be represented in the analytical supply function equilibrium model due to the mathematical problems concerning the calculation of the equilibrium2. The computational model is then applied to analyse different options for the second round of plant divestment in England and Wales in 1999. In the simulation runs the demand slope and the volume of contracts for differences are varied. Their results show that the analysed divestment options result in lower average percentage bids above marginal cost.
But, the authors conclude, the proposed divestiture still leaves market power with generators in the short-term, and could result in prices more than 20 % above short-run marginal costs.
In a later paper the authors (Bunn and Day, 2002) present this model as a competitive bench-mark against which to assess generator conduct and to diagnose the separate causes of bench-market structure and market conduct in situations where prices appear to be above marginal costs.
For the tested scenarios the simulated system supply functions are above the marginal cost function and significantly below the system supply curve observed in the England and Wales pool on an exemplary day, except at low demand levels. This leads the authors to the conclu-sion that the extent to which the simulated supply functions are above the marginal cost tion is caused by the market structure. Based on the fact that the observed system supply func-tions in the real-world market are still above the simulated system supply funcfunc-tions, Bunn and Day conclude that there is collusion within the market, thus identifying a problem of market conduct.
Bower and Bunn present an agent-based simulation of the England and Wales electricity mar-ket in the year 2000 (Bower and Bunn, 2000). The simulation is designed to compare differ-ent market mechanisms. Examples are the comparison of daily bids versus hourly bids and the comparison of uniform price and discriminatory settlement. In the given simulation, genera-tion agents bid on a single electricity market. The electricity demand is modelled by a price inelastic aggregate demand curve. Agents are endowed with a simple reinforcement learning algorithm which is driven by the goal to maximize profits and to reach a given utilization rate of a power plant. The agents can choose between four strategies, e. g. lowering prices when the expected utilization is not met. Since bids are calculated for every plant, generation com-panies with more plants get more insights into the market. The memory of the agents is lim-ited to two days. A more detailed discussion of the results of an application of this model can be found in (Bower and Bunn, 2001). The results of the agent-based simulation show that the discriminatory settlement where the individual bid price is paid for successful bids leads to higher prices than uniform price settlement where the marginal bid sets the market price for
2 see also Chapter 3.3.2.2
all bids. Another finding is that a shift from daily bidding with a single price for the entire day to hourly bidding leads to higher prices which can be explained by the fact that the inelastic demand helps the generation companies to reach high market prices in peak demand periods.
In a next step the simulation results are compared to classical economic models of monopoly, duopoly and perfect competition supporting the simulation outcomes with regard to the com-parison of uniform and discriminatory settlement.
In (Bower et al., 2001) an application of the developed model for the simulation of the Ger-man electricity sector is described. Thereby the GerGer-man market is simulated as a day-ahead bilateral market with simultaneous bidding and pay as bid market clearing. The demand is represented by an aggregate demand curve. Transmission constraints or costs are neglected. In the given case study the impact of the mergers between RWE/VEW and Preussen Elec-tra/Bayernwerk creating Germany's biggest electricity utilities is analysed in different simula-tion settings including plant closures and lower utilizasimula-tion targets for power plants. In these scenarios electricity prices rise considerably as an effect of the mergers.
An extension of the agent-based simulation is described by Bunn and Oliveira (Bunn and Oliveira, 2001). The given paper explicitly models electricity demand as actively bidding sup-plier agents. Supply-agents are characterized by a given market share, contract cover, a given retail price, the mean prediction error in forecasting the contracted load and a search propen-sity which indicates the agents' willingness to search for new strategies. The supplier agents are driven by the goal to maximize daily profits while keeping balancing market exposure caused by insufficient contract cover close to a target value. Another enhancement of the model is a more detailed simulation of the supply-side by integrating plant cycles and avail-ability into the generator characteristics. Other aspects like contract cover, search propensity and goals are similar to the supply agents. The extended model incorporates the balancing power market as an additional market. Both markets are modelled as sequential markets.
Trading takes place on a day-ahead basis for single hours. The balancing power market is executed after the power exchange generators and suppliers are allowed to bid on the balanc-ing markets in order to correct load prediction or missbalanc-ing contract cover.
In a case study Bunn and Oliveira (Bunn and Oliveira, 2003) use the developed model in or-der to analyse whether two generation companies on the UK electricity market are capable of increasing their profits by manipulating market prices. They can show that in a one shot Be-trand game3, where price is the strategic decision variable, generators are capable of reaching prices above marginal cost. Unilateral capacity withholding leads to increased profits. Based on these results, Bunn and Oliveira argue that profit margins are not a good indicator to evaluate market power abuse and they state that learning in repeated games has to be taken into account. In order to reach more realistic results, the developed agent-based simulation is applied for an analysis of the England and Wales electricity market in the year 2000. The simulation is carried out for six typical demand profiles and six predefined strategies for the two generators. Results of the simulation runs indicate that only one generator is capable of increasing power exchange prices unilaterally. In order to manipulate prices profitably, both
3 see also Chapter 3.3.2.2 for more information on Betrand models
generation companies have to act together. However, prices on the balancing market seem to be more robust against manipulation from both players. Based on these results, the authors argue that additional insights can be gained in repeated games using agent learning algorithms to explore phenomena like implicit collusion.
Recently the described model has been applied to the analysis of market power on the elec-tricity market in England and Wales (Bunn and Martoccia, 2005). A new direction of research seeks to extend the developed simulation platform for the analysis of the impact of crosshold-ings and vertical integration. Thereby the analysis of the gas to power value chain is within the centre of research. A new challenge is the simulation of two markets for different com-modities (Micola et al., 2006).